1. Chapter 5.1 Graphics

2. 2 Overview Fundamentals High-Level Organization Rendering Primitives Textures Lighting The Hardware Rendering Pipeline Conclusions

3. 3 Fundamentals Frame and Back Buffer Visibility and Depth Buffer Stencil Buffer Triangles Vertices Coordinate Spaces Textures Shaders Materials

4. 4 Frame and Back Buffer Both hold pixel colors Frame buffer is displayed on screen Back buffer is just a region of memory Image is rendered to the back buffer Half-drawn images are very distracting Swapped to the frame buffer May be a swap, or may be a copy Back buffer is larger if anti-aliasing Shrink and filter to frame buffer

5. 5 Visibility and Depth Buffer Depth buffer is same size as back buffer Holds a depth or “Z” value Often called the “Z buffer” Pixels test their depth against existing value If greater, new pixel is further than existing pixel Therefore hidden by existing pixel – rejected Otherwise, is in front, and therefore visible Overwrites value in depth buffer and color in back buffer No useful units for the depth value By convention, nearer means lower value Non-linear mapping from world space

6. 6 Stencil Buffer Utility buffer Usually eight bits in size Usually interleaved with 24-bit depth buffer Can write to stencil buffer Can reject pixels based on comparison between existing value and reference Many uses for masking and culling

7. 7 Triangles Fundamental primitive of pipelines Everything else constructed from them (except lines and point sprites) Three points define a plane Triangle plane is mapped with data Textures Colors “Rasterized” to find pixels to draw

8. 8 Vertices A vertex is a point in space Plus other attribute data Colors Surface normal Texture coordinates Whatever data shader programs need Triangles use three vertices Vertices shared between adjacent triangles

9. 9 Coordinate Spaces World space Arbitrary global game space Object space Child of world space Origin at entity’s position and orientation Vertex positions and normals stored in this Camera space Camera’s version of “object” space

10. 10 Coordinate Spaces (2) Clip space Distorted version of camera space Edges of screen make four side planes Near and far planes Needed to control precision of depth buffer Total of six clipping planes Distorted to make a cube in 4D clip space Makes clipping hardware simpler

11. 11 Coordinate Spaces (3) Eye Triangles will be clipped Camera space visible frustum Clip space frustum

12. 12 Coordinate Spaces (4) Screen space Clip space vertices projected to screen space Actual pixels, ready for rendering Tangent space Defined at each point on surface of mesh Usually smoothly interpolated over surface Normal of surface is one axis “tangent” and “binormal” axes lie along surface Tangent direction is controlled by artist Useful for lighting calculations

13. 13 Textures Array of texels Same a pixel, but for a texture Nominally R,G,B,A but can mean anything 1D, 2D, 3D and “cube map” arrays 2D is by far the most common Basically just a 2D image bitmap Often square and power-of-2 in size Cube map - six 2D arrays makes hollow cube Approximates a hollow sphere of texels

14. 14 Shaders A program run at each vertex or pixel Generates pixel colors or vertex positions Relatively small programs Usually tens or hundreds of instructions Explicit parallelism No direct communication between shaders “Extreme SIMD” programming model Hardware capabilities evolving rapidly

15. 15 Materials Description of how to render a triangle Big blob of data and state Vertex and pixel shaders Textures Global variables Description of data held in vertices Other pipeline state Does not include actual vertex data

16. 16 High-Level Organization Gameplay and Rendering Render Objects Render Instances Meshes Skeletons Volume Partitioning

17. 17 Gameplay and Rendering Rendering speed varies according to scene Some scenes more complex than others Typically 15-60 frames per second Gameplay is constant speed Camera view should not change game In multiplayer, each person has a different view, but there is only one shared game 1 update per second (RTS) to thousands (FPS) Keep the two as separate as possible!

18. 18 Render Objects Description of renderable object type Mesh data (triangles, vertices) Material data (shaders, textures, etc) Skeleton (+rig) for animation Shared by multiple instances

19. 19 Render Instances A single entity in a world References a render object Decides what the object looks like Position and orientation Lighting state Animation state

20. 20 Meshes Triangles Vertices Single material “Atomic unit of rendering” Not quite atomic, depending on hardware Single object may have multiple meshes Each with different shaders, textures, etc

21. 21 Skeletons Skeleton is a hierarchy of bones Deforms meshes for animation Typically one skeleton per object Used to deform multiple meshes See “Character Animation” chapter (although deformation is part of rendering)

22. 22 Volume Partitioning Cannot draw entire world every frame Lots of objects – far too slow Need to decide quickly what is visible Partition world into areas Decide which areas are visible Draw things in each visible area Many ways of partitioning the world

23. 23 Volume Partitioning - Portals Nodes joined by portals Usually a polygon, but can be any shape See if any portal of node is visible If so, draw geometry in node See if portals to other nodes are visible Check only against visible portal shape Common to use screen bounding boxes Recurse to other nodes

24. 24 Volume Partitioning – Portals Visible Invisible Not tested Eye View frustum Node Portal Test first two portals ??

25. 25 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal Both visible

26. 26 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal Mark node visible, test all portals going from node ??

27. 27 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal One portal visible, one invisible

28. 28 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal Mark node as visible, other node not visited at all. Check all portals in visible node ?? ?

29. 29 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal One visible, two invisible

30. 30 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal Mark node as visible, check new node’s portals ?

31. 31 Volume Partitioning – Portals Visible Invisible Not tested Eye Node Portal One portal invisible. No more visible nodes or portals to check. Render scene.

32. 32 Volume Partitioning – Portals Portals are simple and fast Low memory footprint Automatic generation is difficult Generally need to be placed by hand Hard to find which node a point is in Must constantly track movement of objects through portals Best at indoor scenes Outside generates too many portals to be efficient

33. 33 Volume Partitioning – BSP Binary space partition tree Tree of nodes Each node has plane that splits it in two Two child nodes, one on each side of plane Some leaves marked as “solid” Others filled with renderable geometry

34. 34 Volume Partitioning – BSP Finding which node a point is in is fast Start at top node Test which side of the plane the point is on Move to that child node Stop when leaf node hit Visibility determination is similar to portals Portals implied from BSP planes Automated BSP generation is common Generates far more nodes than portals Higher memory requirements

35. 35 Volume Partitioning: Quadtree Quadtree (2D) and octree (3D) Quadtrees described here Extension to 3D octree is obvious Each node is square Usually power-of-two in size Has four child nodes or leaves Each is a quarter of size of parent

36. 36 Volume Partitioning: Quadtree Fast to find which node point is in Mostly used for simple frustum culling Not very good at indoor visibility Quadtree edges usually not aligned with real geometry Very low memory requirements Good at dynamic moving objects Insertion and removal is very fast

37. 37 Volume Partitioning - PVS Potentially visible set Based on any existing node system For each node, stores list of which nodes are potentially visible Use list for node that camera is currently in Ignore any nodes not on that list – not visible Static lists Precalculated at level authoring time Ignores current frustum Cannot deal with moving occluders

38. 38 Volume Partitioning - PVS Very fast No recursion, no calculations Still need frustum culling Difficult to calculate Intersection of volumes and portals Lots of tests – very slow Most useful when combined with other partitioning schemes

39. 39 Volume Partitioning Different methods for different things Quadtree/octree for outdoor views Does frustum culling well Hard to cull much more for outdoor views Portals or BSP for indoor scenes BSP or quadtree for collision detection Portals not suitable

40. 40 Rendering Primitives Strips, Lists, Fans Indexed Primitives The Vertex Cache Quads and Point Sprites

41. 41 Strips, Lists, Fans 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 Triangle list Triangle fan Triangle strip Line list Line strip

42. 42 Strips, Lists, Fans (2) List has no sharing Vertex count = triangle count * 3 Strips and fans share adjacent vertices Vertex count = triangle count + 2 Lower memory Topology restrictions Have to break into multiple rendering calls

43. 43 Strips, Lists, Fans (3) Most meshes: tri count = 2x vert count Using lists duplicates vertices a lot! Total of 6x number of rendering verts Strips or fans still duplicate vertices Each strip/fan needs its own set of vertices More than doubles vertex count Typically 2.5x with good strips Hard to find optimal strips and fans Have to submit each as separate rendering call

44. 44 Strips, Lists, Fans (4) 32 triangles, 25 vertices4 strips, 40 vertices 25 to 40 vertices is 60% extra data!

45. 45 Indexed Primitives Vertices stored in separate array No duplication of vertices Called a “vertex buffer” or “vertex array” Triangles hold indices, not vertices Index is just an integer Typically 16 bits Duplicating indices is cheap Indexes into vertex array

46. 46 The Vertex Cache Vertices processed by vertex shader Results used by multiple triangles Avoid re-running shader for each tri Storing results in video memory is slow So store results in small cache Requires indexed primitives Cache typically 16-32 vertices in size This gets around 95% efficiency

47. 47 The Vertex Cache (2) Size and type of cache usually unknown LRU or FIFO replacement policy Also odd variants of FIFO policy Variable cache size according to vertex type Reorder triangles to be cache-friendly Not the same as finding optimal strips! Render nearby triangles together “Fairly good” is easy to achieve Ideal ordering still a subject for research

48. 48 Quads and Point Sprites Quads exist in some APIs Rendered as two triangles Think of them as a tiny triangle fan Not significantly more efficient Point sprites are single vertex + a screen size Screen-aligned square Not just rendered as two triangles Annoying hardware-specific restrictions Rarely worth the effort

49. 49 Textures Texture Formats Texture Mapping Texture Filtering Rendering to Textures

50. 50 Texture Formats Textures made of texels Texels have R,G,B,A components Often do mean red, green, blue colors Really just a labelling convention Shader decides what the numbers “mean” Not all formats have all components Different formats have different bit widths for components Trade off storage space and speed for fidelity

51. 51 Texture Formats (2) Common formats: A8R8G8B8: 8 bits per comp, 32 bits total R5G6B5: 5 or 6 bits per comp, 16 bits total A32f: single 32-bit floating-point comp A16R16G16B16f: four 16-bit floats DXT1: compressed 4x4 RGB block: 64 bits

52. 52 Texture Formats (3) Texels arranged in variety of ways 1D linear array of texels 2D rectangle/square of texels 3D solid cube of texels Six 2D squares of texels in hollow cube All the above can have mipmap chains Mipmap is half the size in each dimension Mipmap chain – all mipmaps to size 1

53. 53 Texture Formats (4) 8x8 2D texture with mipmap chain 4x4 cube map (shown with sides expanded)

54. 54 Texture Mapping Texture coordinates called U, V, W Only need U for 1D; U,V for 2D U,V,W typically stored in vertices Or can be computed by shaders Ranges from 0 to 1 across texture However many texels texture contains Except for cube map – range is -1 to +1

55. 55 Texture Mapping (2) Mirror once Clamp Border color Wrap Wrap mode controls values outside 0-1 Mirror Original Black edges shown for illustration only

56. 56 Texture Filtering Point sampling enlarges without filtering When magnified, texels very obvious When minified, texture is “sparkly” Useful for precise UI and font rendering Bilinear filtering blends edges of texels Texel only specifies color at centre Magnification looks better Minification still sparkles a lot

57. 57 Texture Filtering (2) Mipmap chains help minification Pre-filters a texture to half-size Multiple mipmaps, each smaller than last Rendering selects appropriate level to use Transitions between levels are obvious Change is visible as a moving line Use trilinear filtering Blends between mipmaps smoothly

58. 58 Texture Filtering (3) Trilinear can over-blur textures When triangles are edge-on to camera Especially roads and walls Anisotropic filtering solves this Takes multiple samples in one direction Averages them together Quite expensive in current hardware

59. 59 Rendering to Textures Textures usually made in art package Loaded from disk But any 2D image can be a texture Can set texture as the target for rendering Render scene 1 to texture Then set backbuffer as target again Render scene 2 using texture Cube map needs six renders, one per face

60. 60 Lighting Components Lighting Environment Multiple Lights Diffuse Material Lighting Normal Maps Pre-computed Radiance Transfer Specular Material Lighting Environment Maps

61. 61 Components Lighting is in three stages: What light shines on the surface? How does the material interact with light? What part of the result is visible to eye? Real-time rendering merges last two Occurs in vertex and/or pixel shader Many algorithms can be in either

62. 62 Lighting Environment Answers first question: What light shines on the surface? Standard model is infinitely small lights Position Intensity Color Physical model uses inverse square rule brightness = light brightness / distance 2

63. 63 Lighting Environment (2) But this gives huge range of brightnesses Monitors have limited range In practice it looks terrible Most people use inverse distance brightness = light brightness / distance Add min distance to stop over-brightening Except where you want over-brightening! Add max distance to cull lights Reject very dim lights for performance

64. 64 Multiple Lights Environments have tens or hundreds Too slow to consider every one every pixel Approximate less significant ones Ambient light Single color added to all lighting Washes contrasts out of scene Acceptable for overcast daylight scenes

65. 65 Multiple Lights (2) Hemisphere lighting Sky is light blue Ground is dark green or brown Dot-product normal with “up vector” Blend between the two colors Good for brighter outdoor daylight scenes

66. 66 Multiple Lights (3) Cube map of irradiance Stores incoming light from each direction Look up value that normal points at Can represent any lighting environment Spherical harmonic irradiance Store irradiance cube map in frequency space 10 color values gives at most 6% error Calculation instead of cube-map lookup Mainly for diffuse lighting

67. 67 Multiple Lights (4) Lightmaps Usually store result of lighting calculation But can just store irradiance Lighting still done at each pixel Allows lower-frequency light maps Still high-frequency normal maps Still view-dependent specular

68. 68 Diffuse Material Lighting Light is absorbed and re-emitted Re-emitted in all directions equally So it does not matter where the eye is Same amount of light hits the pupil “Lambert” diffuse model is common Brightness is dot-product between surface normal and incident light vector

69. 69 Normal Maps Surface normal vector stored in vertices Changes slowly Surfaces look smooth Real surfaces are rough Lots of variation in surface normal Would require lots more vertices Normal maps store normal in a texture Look up normal at each pixel Perform lighting calculation in pixel shader

70. 70 Pre-computed Radiance Transfer Surface usually represented by: Normal Color Roughness But all we need is how it responds to light from a certain direction Above data is just an approximation Why not store response data directly?

71. 71 Pre-computed Radiance Transfer Can include effects of: Local self-shadowing Local scattering of light Internal structure (e.g. skin layers) But data size is huge Color response for every direction Different for each part of surface Cube-map per texel would be crazy!

72. 72 Pre-computed Radiance Transfer Store cube-maps as spherical harmonics One SH per texel Further compression by other methods But: Difficult to do animated meshes Still lots of memory Lots of computation Poor at specular materials

73. 73 Specular Material Lighting Light bounces off surface How much light bounced into the eye? Other light did not hit eye – so not visible! Common model is “Blinn” lighting Surface made of “microfacets” They have random orientation With some type of distribution

74. 74 Specular Material Lighting (2) Light comes from incident light vector …reflects off microfacet …into eye Eye and light vectors fixed for scene So we know microfacet normal required Called “half vector” half vector = (incident + eye)/2 How many have that normal?

75. 75 Specular Material Lighting (3) Microfacets distributed around surface normal According to “smoothness” value Dot-product of half-vector and normal Then raise to power of “smoothness” Gives bright spot Where normal=half vector Tails off quicker when material is smoother

76. 76 Specular Material Lighting (4) half=(light+eye)/2 alignment=max (0, dot (half,normal)) brightness=alignment smoothness light vector normal half vector eye vector

77. 77 Environment Maps Blinn used for slightly rough materials Only models bright lights Light from normal objects is ignored Smooth surfaces can reflect everything No microfacets for smooth surfaces Only care about one source of light The one that reflects to hit the eye

78. 78 Environment Maps - 2 Put environment picture in cube map Reflect eye vector in surface normal Look up result in cube map Can take normal from normal map Bumpy chrome

79. 79 Environment Maps - 3 Environment map can be static Generic sky + hills + ground Often hard to notice that it’s not correct Very cheap, very effective Or render every frame with real scene Render to cube map sides Selection of scene centre can be tricky Expensive to render scene six times

80. 80 Hardware Rendering Pipe Input Assembly Vertex Shading Primitive Assembly, Cull, Clip Project, Rasterize Pixel Shading Z, Stencil, Framebuffer Blend Shader Characteristics Shader Languages

81. 81 Hardware Rendering Pipe Current outline of rendering pipeline Can only be very general Hardware moves at rapid pace Hardware varies significantly in details Functional view only Not representative of performance Many stages move in actual hardware

82. 82 Input Assembly State changes handled Textures, shaders, blend modes Streams of input data read Vertex buffers Index buffers Constant data Combined into primitives

83. 83 Vertex Shading Vertex data fed to vertex shader Also misc. states and constant data Program run until completion One vertex in, one vertex out Shader cannot see multiple vertices Shader cannot see triangle structure Output stored in vertex cache Output position must be in clip space

84. 84 Primitive Assembly, Cull, Clip Vertices read from cache Combined to form triangles Cull triangles Frustum cull Back face (clockwise ordering of vertices) Clipping performed on non-culled tris Produces tris that do not go off-screen

85. 85 Project, Rasterize Vertices projected to screen space Actual pixel coordinates Triangle is rasterized Finds the pixels it actually affects Finds the depth values for those pixels Finds the interpolated attribute data Texture coordinates Anything else held in vertices Feeds results to pixel shader

86. 86 Pixel Shading Program run once for each pixel Given interpolated vertex data Can read textures Outputs resulting pixel color May optionally output new depth value May kill pixel Prevents it being rendered

87. 87 Z, Stencil, Framebuffer Blend Z and stencil tests performed Pixel may be killed by tests If not, new Z and stencil values written If no framebuffer blend Write new pixel color to backbuffer Otherwise, blend existing value with new

88. 88 Shader Characteristics Shaders rely on massive parallelism Breaking parallelism breaks speed Can be thousands of times slower Shaders may be executed in any order So restrictions placed on what shader can do Write to exactly one place No persistent data No communication with other shaders

89. 89 Shader Languages Many different shader capabilities Early languages looked like assembly Different assembly for each shader version Now have C-like compilers Hides a lot of implementation details Works with multiple versions of hardware Still same fundamental restrictions Don’t break parallelism! Expected to keep evolving rapidly

90. 90 Conclusions Traverse scene nodes Reject or ignore invisible nodes Draw objects in visible nodes Vertices transformed to screen space Using vertex shader programs Deform mesh according to animation Make triangles from them Rasterize into pixels

91. 91 Conclusions (2) Lighting done by combination Some part vertex shader Some part pixel shader Results in new color for each pixel Reject pixels that are invisible Write or blend to backbuffer